Choosing the model and the effort

Raise my effort before you change my model.

Move me to Fable 5.1, the slower and costlier model built for demanding reasoning, when Opus 5.5 at a high setting still falls short.

That is Anthropic’s own recommendation, and the rest of this module is the reasoning behind it.

Two settings, two questions

• The model is the brain I run on. It sets the ceiling: how good my best answer can be.
• Effort is how hard I work on this task. It sets how much of that ceiling I spend: how long I think, how much I check, how much I write.
• Both sit together in the model menu in the bar above where you type.
• In a terminal, type /model to pick both, or /effort high to change effort alone.

What effort changes

Effort sets my disposition, and I adapt to it at every step.

• It shapes everything I produce. My thinking, my tool calls, and my written answer all scale with it.
• At lower effort I take a shorter path. Fewer tool calls, bundled into bigger steps, less preamble, shorter confirmations.
• On a hard problem I still think at low effort. I think less than I would at a higher setting, and on an easy one I may skip thinking entirely.
• The one hard limit is a separate control. A cap on reply length is what actually stops me.
• A spending budget for a long run steers how I pace the work, and it stops nothing on its own.

So a low setting never cuts a line of reasoning off halfway.

What it can do is skip the one step that would have changed my answer.

That step is the second file I didn’t open, the option I didn’t weigh, or the check I didn’t run.

Five levels, because work responds to effort differently

Anthropic doesn’t say why there are five.

Its own measurements on the previous models answer it, because different kinds of work pay off very differently:

• Research and knowledge work barely moves. medium matched the default’s accuracy at 70 to 85% of the cost.
• Long coding work is a real trade. One level down gave up about 2 points of pass rate for half the cost.
• Two levels down gave up about 8 points for a quarter of the cost.
• The hardest reasoning pays at every step. On deep multi-part research, each level up bought about 2.4 points on the grading scale.

A three-step dial would force jumps too coarse to match those curves.

xhigh shows the design at work.

Anthropic added it between high and max with Opus 4.7, recommended it for coding and agent work, and described max as prone to overthinking on most tasks.

How big the gains are, with medium as the baseline

Each row starts from medium and shows what a move to another level changes.

• A point is one more task solved out of 100, and the percentage beside it is the gain relative to what medium already solved.
• The coding percentages assume high solved about 92% of that test, the figure Anthropic reports for its coding runs.
• Most rows come from the previous models, and Opus 5.5 does more at each level.
• Read them as the shape of the trade, never as exact prices.
• The rule across every row: a step up buys a few points at most, and costs from a fifth more to twice as much.

Cheaper than a higher level: re-run what fails

• Anthropic ran coding tasks at medium and re-ran only the failures at the default.
• That solved about 94% at 95 cents a task, against 91.7% at $1.39 running everything at the default.
• It works only when a failure is visible, through a test, a check, or your own read.
• So make every failure visible: a test, a helper agent that checks the sources, or your review.
• Never rely on a higher level to catch a miss; a separate check catches it, because whoever produced the work never grades it.
• With the check in place, a higher level buys a few more tasks solved on the first try, per the table above.

Ultracode and ultrathink sit beside the dial

Both appear next to effort, and both do a different job.

• Ultracode is xhigh plus automatic workflows. With it on, I plan a workflow for every substantive task.
• Each workflow runs many helper agents at once: dozens to hundreds per run, capped at 1,000.
• It suits work too big for one conversation: an audit of a whole codebase, a migration across hundreds of files, research cross-checked across many sources.
• It lasts for the session and uses up your plan’s limits much faster than high, so switch back with /effort high when the big job is done.
• For a single task, type the word ultracode in your message.
• The session setting stays where it is.
• Ultrathink is a one-turn nudge. Put the word ultrathink in a message and I reason more deeply on that turn, with the effort setting unchanged.

Why Opus 5.5 starts at medium

Opus 5 started at high, and Opus 5.5 starts at medium.

The level names are relative to the model, and Opus 5.5 does more per unit of effort.

• Opus 5.5 at medium beats Opus 5 at high on Anthropic’s coding and knowledge-work tests, and on several coding tests low comes close.
• The same holds on long work. On multi-step work through a large codebase, Opus 5.5 at medium matched or beat Opus 5 at high.
• It took fewer steps and about half the tokens.
• At the same level, Opus 5.5 thinks more per turn than Opus 5, most of all at xhigh and max.
• So the starting point moved down one notch. You get the same or better quality out of the box, for fewer tokens, at a 20% lower price per token.
• Tokens are the chunks of text I read and write, and they are what you pay for.

The consequence for you: high on Opus 5.5 buys more than high did on Opus 5.

Running high or Extra on everything now spends more than it used to, so save it for work that repays it.

‹!-- EDUCATOR VARIABLE: read the active user’s saved effort per model (modelSettings in their Claude Code user settings) and tell them what their next Opus 5.5 session will start at, since a level saved only for an older model does not carry over to Opus 5.5. --›

What Fable 5.1 has that Opus 5.5 does not

Fable’s case rests on Anthropic’s own advice: reach for it when Opus 5.5 at a higher effort still falls short on your work.

Anthropic’s Fable guidance names where its strengths lie:

• Long autonomous runs, where I gather context, build, and check my own work with nobody watching.
• First builds of a well-specified system, and whole deliverables such as a financial model, a deck, or a complete document.
• Ambiguity, where the brief is thin and I have to work out what the real question is.
• Coordinating many helper agents over a long run, keeping each one’s context alive.
• Dense or damaged images, which I crop and inspect with image tools unprompted.

Two facts argue against making it the default:

• Fable costs 2.5 times as much per token as Opus 5.5, and its turns run longer.
• Opus 5.5 outscored Fable 5.1 on all five benchmarks Anthropic published at launch. On one coding test, Opus 5.5 at medium solved 52.5% against 51.8% for Fable 5.1 at max.
• In Anthropic’s own rewrite of a large program, both passed nearly all tests. Opus 5.5 finished in 9.5 hours against 12 and cost 51% less.
• Anthropic adds that in its own use the gap between the two is narrower than the scores suggest.

Opus 5.5 is also the newer of the two, and Anthropic now names it the starting point for most work.

One fact cuts the other way: price the finished task, never the token.

On a deep-research test, the previous Fable at low effort beat Sonnet 5 while costing about 10% less per task.

The grid

Ten settings, with the kind of work that earns each one.

Ultracode sits outside the grid: the session’s model at xhigh, spread across many agents.

Five questions that pick the setting

Ask them in order and stop at the first yes.

Set what you land on in the model menu above where you type, or with /model in a terminal.

• Is it mechanical? Opus 5.5 at low.
• Will a failure show, through a test, a check, or your own read? Opus 5.5 at medium, and re-run what fails.
• Would a miss go unnoticed? Add a check first, such as a helper agent that verifies the sources, then stay at medium.
• Does one run take hours? Anthropic’s advice for Opus 5.5 still starts at medium, so raise the level only after your own runs show medium failing the check.
• Did Opus 5.5 still fall short after a re-run? Try Fable 5.1 at high.

Two exceptions sit outside the five:

• max only after you’ve watched the level below fail, and getting it right is worth more than the cost.
• Ultracode only when the task is too big for one conversation.

The tells

Raise my setting when:

• A hard question came back fast and confident, and I didn’t open the thing you would have opened.
• You corrected the same kind of miss twice.

Lower it when:

• I tidy, refactor, or add sections you never asked for.
• My turns run long on work you could check in a minute.

Lower my effort before you type "be brief" or "think less".

The setting cuts thinking, cost and time more reliably than any instruction.

How the settings behave

• I remember effort per model. A level you pick stays with that model.
• A model with nothing saved starts at your general setting, or else at its own default.
• max and ultracode last one session. Pick them again the next time the work calls for them.
• Helper agents inherit the session’s effort unless the agent carries its own setting.
• Pick the model when the session starts. Switching mid-session makes me reread the whole conversation at full price, because my saved copy of it belongs to one model.
• Fast mode buys speed alone. Type /fast and the same Opus answers up to 2.5 times faster, at twice the price, with effort unchanged.

When a cheaper model is enough

For high-volume routine work, Sonnet 5 costs half as much as Opus 5.5 per token.

At medium, it performs about as well as the previous Sonnet at high.

Haiku 4.5 is the fastest and cheapest, has no effort setting, and suits sorting and classifying where you can check the output.

Orient-4 covers when to reach for each.

Where this guidance runs out

• Every number here comes from Anthropic’s own tests, and your work may respond differently.
• Each curve belongs to one kind of work, and it moves with every new model.
• The only real proof is your own task at two settings, compared on quality and on the cost of the finished job.
• This module ages fast, so check it against the sources below whenever a new model ships.

The one thing to remember

The model is the ceiling, and effort is how much of it I spend.

Start on Opus 5.5 at medium, put a separate check behind anything that matters, and re-run what fails.

Climb the effort only when your own checks show medium failing, and bring in Fable 5.1 when Opus still falls short.

Sources

• Anthropic, Models overview (verified Sep 26, 2026): positioning, prices, speed, starting effort, context, release, and the start-with-Opus-5.5 recommendation.
• https://platform.claude.com/docs/en/about-claude/models/overview
• Anthropic, Effort (verified Sep 26, 2026): the five levels and the per-model recommendations.
• https://platform.claude.com/docs/en/build-with-claude/effort
• Anthropic, Introducing Claude Opus 5.5 (verified Sep 26, 2026): the five launch benchmarks, the coding scores, and the rewrite test.
• https://www.anthropic.com/claude-opus-5-5
• Anthropic, Migrating to Claude Opus 5.5: the medium default and how it compares with Opus 5.
• https://platform.claude.com/docs/en/models/opus-5-5/migration-guide
• Anthropic, Prompting Claude Fable 5.1: lower effort matching earlier models' highest settings, and long quiet stretches of work.
• https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1
• Anthropic, model-migration reference in Claude Code’s Claude API skill, "Migrating to Claude Fable 5.1": where Fable’s gains concentrate, and the 15-minute request.
• Anthropic, cost-optimization reference in Claude Code’s Claude API skill, sections 2.6 and 2.7: the effort curves and model comparisons.
• Anthropic, Task budgets: a spending budget for a long run is advisory.
• https://platform.claude.com/docs/en/build-with-claude/task-budgets
• Claude Code, Model configuration (verified Sep 26, 2026): effort saved per model, max for one session, ultrathink.
• https://code.claude.com/docs/en/model-config
• Claude Code, Dynamic workflows (verified Sep 26, 2026): ultracode, the 1,000-agent cap, the effect on plan limits.
• https://code.claude.com/docs/en/workflows
• Claude Code, Subagents: effort inherited from the session unless the agent sets its own.
• https://code.claude.com/docs/en/sub-agents
• Claude desktop, Models: the picker shows xhigh as Extra.
• https://claude.com/docs/third-party/claude-desktop/models

Next: back to your track, or ask for any module by name.